Research On Low Light Image Enhancement Based On Retinex Theory And Deep Learning

Due to the limitations of environmental factors and camera equipment,especially at night,the images captured will appear low brightness,low contrast,high noise and color distortion,which affects the subjective visual perception of human images.In addition,in advanced computer vision processing tasks,a large amount of image feature information in low-illumination images is not obvious,which makes subsequent tasks such as image recognition,target detection and target tracking difficult.In order to improve the overall quality of low-light images and maintain the details of images while enhancing brightness,color and contrast,this paper proposes the following innovations:(1)In order to solve the problem that the traditional method based on Retinex theory uses Gaussian filtering or sets specific prior constraints to estimate the image illuminance component,which leads to inaccurate illumination estimation,resulting in halo artifacts and color distortion,this paper transforms the low-illuminance image into HSV space,and uses the iterative multi-scale guided filtering algorithm to accurately estimate the illuminance component of the V-component weight.Then,the illuminance component is corrected by using two-dimensional adaptive gamma correction and hierarchical CLAHE algorithm.Since low illumination images often contain noise,this paper adopts the method of combined Gaussian filtering and guided filtering to suppress the noise in the reflection component.Finally,the enhanced V component is obtained by multiplying the illuminance component and reflection component element by element,and the enhanced image is converted back to RGB space to obtain the final enhanced result.(2)In order to solve the problems of image color distortion,insufficient brightness and noise amplification in deep learning methods based on Retinex theory,this paper proposes a low-illumination image enhancement network based on Retinex theory and attention mechanism.The network consists of decomposition network,reflection denoising network and illumination adjustment network.The decomposition network adopts two-branch network structure,and adopts channel splicing to realize skip connection between the upper and lower sampling layers,so as to improve image resolution and make up for the details lost in the process of downsampling,and accurately extract the illuminance and reflection components from the input image.For reflection denoising network,the attention mechanism of convolutional module is introduced to improve the sensitivity of the network to image noise information and suppress the noise in the image.For the illumination adjustment network,the global attention mechanism is introduced to improve the brightness and contrast of the illumination component,and channel stitching is used to correlate the features of up and down sampling to compensate for the details of the illumination component and improve the Global illumination consistency of the enhanced image.